Comparative evaluation of penalized regression models with multiple linear regression for predicting rapeseed-mustard yield: Weather-indices based approach

Abstract

Rapeseed-mustard (Brassica spp.) is one of the important edible oilseeds crops in India. The same level of weather condition impacts the growth and establishment of rapeseed-mustard plant differently in different stages of crop which lead to large intra-seasonal yield variations. Hence it is essential to give weightage to weekly weather conditions while fitting predictive model. In this present study, path-coefficient based weighted index was proposed along with existing unweighted and correlation based weighted index. The performance of penalized regression models viz. Ridge Regression, Least Absolute Shrinkage and Selection Operator (LASSO) and Elastic Net (ENET) were compared with Multiple Linear Regression (MLR) for predicting rapeseed-mustard yield using weather-indices. The results revealed that the path-coefficient based weighting of weather parameters to the yield were stable than correlation based weighted-indices. Path-coefficient based weighted indices of maximum temperature, minimum temperature and windspeed were important variables in projection of yield. The performance of MLR was poor during validation of model due to overfitting issue. The performance of penalized models was stable in both calibration and validation of the model. The LASSO and ENET models that accompanied with coefficient shrinkage and variable selection were found to be the best fitted models for predicting Rapeseed-Mustard yield.